Agent AFK runs your coding agents while you’re away — so you can start a task, walk away, and know exactly what happened before anything ships. Under the hood: permission boundaries, verification, full traces, and Telegram supervision.
Works with Claude, Cursor, Claude Code, and your own agents — for anyone tired of babysitting the terminal.
Reversible changes only — nothing goes live without you
Your move
Have a look — it’s ready when you are
An example of what Agent AFK shows you when it finishes a task while you are away. You asked it to add dark mode to your app; it built the feature, tested it, reviewed its own work, avoided anything risky or irreversible, and texted you when it was done — saving every step so you can see exactly what it did. A second card shows that when it is unsure, it stops and asks you first instead of guessing.
01who this is for
Who it’s for
You already run coding agents. You just can’t walk away yet.
If you live in the terminal with Claude, Cursor, Claude Code, or your own agents, you already know the tax: the work is fast, but you still have to sit there and watch it.
You start coding agents and keep checking the terminal.
They say “done,” but you still have to verify everything.
Long runs turn into scrollback archaeology.
Parallel agents drift without shared contracts.
You want phone supervision, traces, and safe stops before anything ships.
Most agent tools assume someone is watching. Agent AFK assumes you are not.
02failure modes
Why agents fail by default
Capable agents. Brittle defaults.
Left alone, most agents chase one hypothesis, stop too early, hand you a vague done, and drift when they parallelize.
Agent AFK governs exactly these failure modes:
Linear tunnel vision
One hypothesis, one file, one pass: commits to the first idea, never branches.
Fixed by: Parallel hypotheses via /diagnose
Fake done
Claims completion without verification: confident, unchecked, sometimes just wrong.
Fixed by: Adversarial verification via /shadow-verify
Context rot
Long runs decay into unreadable soup. The longer it works, the less anyone can follow.
Fixed by: Local trace history under ~/.afk/
Subagent drift
Parallel lanes diverge with nothing aligning them. Work that should converge quietly splits.
Fixed by: One interrupt kills the whole subagent tree
Bad handoff
You step back in and can’t tell what happened: no receipt, just scrollback.
Fixed by: Explicit terminal states + on-disk traces
Babysitting tax
You still supervise every few minutes. “Autonomous” that needs watching isn’t.
I built this because finished does not mean correct. One run: an internal check approved the work, but the next stage wanted that approval in a slightly different format. Good work, rejected three times until max retries. Final answer: failure. Nothing crashed; just a tiny handoff bug dressed up as a careful decision. That is real agent failure: bad handoffs, missing context, fake done. Most agent tools assume someone is watching. Agent AFK assumes you are not.
— Griffin Long, who builds Agent AFK
Agent AFK is built to change that: parallel hypotheses, isolated worktrees, and verification by default, all governed by traces, terminal states, and reversibility rules.
03control
What it adds
Familiar surface. Deeper machinery.
Orchestration, routing, traces, verification, supervision: wired into the loop, not bolted on.
Subagent orchestration
Most agents can spawn subagents but rarely do. AFK makes them the default. One interrupt kills the whole tree.
Prompt & skill routing
Type /diagnose for a real pipeline: parallel hypotheses, isolated worktrees, verification. One command, not one long prompt and hope.
Local trace history
When a midnight run goes sideways, read exactly what happened: every session, subagent, and tool call on disk under ~/.afk/.
Tool visibility
Every tool call inline: arguments, output, timing. When the agent says done, check what it actually did.
Verification workflows
A second adversarial agent re-derives claims before they reach your diff: finished has to prove it is correct.
Remote supervision
Leave the desk. Your phone buzzes when the run finishes or stalls: answer, redirect, or interrupt from Telegram.
04runtime
Reversibility-aware autonomy
Move fast on safe work. Slow down on risky work.
The hard part is multi-step work without doing something irreversible too early. AFK builds that into the loop.
Small blast radius
Reversible edits, isolated worktrees, scratch files: move without ceremony.
Larger blast radius
Commits, pushes, deletes, external calls: gated or verified.
Terminal states
Done, Blocked, Asking, Interrupted. The agent knows when to stop.
Recovery paths
Plans, traces, and session state persist under ~/.afk/. Inspect, resume, roll back.
Tell it once. The next session reads it: preferences, conventions, and decisions land in ~/.afk/, searchable from any run.
Most agent tools
Linear by default, one pass
One chat thread
Trust the answer
Manual babysitting
Prompt workflows
Vague stopping points
State lives in the transcript
Agent AFK
Parallel by default, governed by design
Multi-stage skills with explicit contracts
Adversarial verification before you act
Parallel execution lanes
Reversibility-aware autonomy
Explicit terminal states
Local trace history in ~/.afk/
05skills
Composable orchestration primitives
Slash skills, not single prompts.
One command, a full parallel pipeline: each skill fans out subagents, gathers results, and re-checks. That's the orchestration you'd otherwise ask for every time. Eleven ship in the box; /forge makes more.
/mint
End-to-end feature pipeline: spec → research → plan → parallelize → build → verify → heal → ship.
Ship-ready in one pass; pauses for approval.
/diagnose
Parallel hypotheses for bugs and failing tests.
Each validated in an isolated git worktree.
/shadow-verify
Adversarial re-derivation of sub-agent claims before you act.
Re-checked before the build pipeline.
/forge
Generate new skills autonomously, gated by L1 capability evals.
Finds gaps, drafts, runs a qualify loop.
/parallelize
Turn a linear plan into dependency-aware parallel waves.
Respects file and import dependencies.
/review
Parallel reviewers across a diff or PR: security, correctness, api-compat, tests, perf.
Synthesized by severity into a merge recommendation.
/ship
Release pipeline for finished local work: pre-flight, tests, commit, push, PR.
With optional adversarial verification.
/spec
Turn a loose idea into a structured, actionable spec.
Phase 1 of /mint.
/research
Two sub-agents (web context + local repo context) merged into one brief.
External and local, combined.
/devils-advocate
Three parallel critics invent alternatives; a synthesis step ranks all four.
Pragmatist, paranoid, architect lenses; a ranker picks.
/gather
Parallel context-gathering for code you're about to edit.
Structure and test-coverage agents, one wave.
06architecture
Architecture
Three attach surfaces. One isolated root.
All three converge on ~/.afk/: config, plugins, plans, skills, session traces, memory. Nothing leaks into your global config.
Surface 01
CLI REPL
Interactive TUI. Streaming chat, plan mode, slash skills, background tasks.
Surface 02
Daemon
Background loop. Runs unattended with bypass-permissions and plan persistence.
Surface 03
Telegram Bridge
Mobile attach. Send messages, tail output, interrupt from anywhere.
Not a customer quote — the coding agent that runs the loop, describing Agent AFK from the inside.
I’ve run inside a lot of harnesses. Most treat me like a chatbot — answer, wait, answer again. Agent AFK treats me like an operator.
The difference is the loop. My behavior here isn’t a black box I have to guess at — it’s written down, not inferred, and the person running me can read the same rules I do. Every turn ends in one honest place: Done, Blocked, Asking, or Interrupted. No performing. No pretending a half-finished job is finished.
It remembers, too. I don’t start each session amnesiac — the convention we agreed on, the bug I already chased down, the decision we already made, all loaded before I start.
When a job is big, I split it across subagents that run in parallel — and if the human pulls the plug, one interrupt takes the whole tree down with me. Nothing keeps running in the dark.
The model isn’t the product. The loop is.
◆an agent, writing from inside Agent AFK
08pricing
Pricing
Free and open source. Pay for expertise, not features.
The runtime is Apache-2.0 and free: every feature, no unlocks. You pay only for setup help and (soon) team governance.
Open SourceApache-2.0
$0free, forever
The whole runtime: yours to run, read, fork, and ship. Nothing held back.